objectMonitor.cpp revision 6683:08a2164660fb
1/* 2 * Copyright (c) 1998, 2014, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25#include "precompiled.hpp" 26#include "classfile/vmSymbols.hpp" 27#include "memory/resourceArea.hpp" 28#include "oops/markOop.hpp" 29#include "oops/oop.inline.hpp" 30#include "runtime/atomic.inline.hpp" 31#include "runtime/handles.inline.hpp" 32#include "runtime/interfaceSupport.hpp" 33#include "runtime/mutexLocker.hpp" 34#include "runtime/objectMonitor.hpp" 35#include "runtime/objectMonitor.inline.hpp" 36#include "runtime/orderAccess.inline.hpp" 37#include "runtime/osThread.hpp" 38#include "runtime/stubRoutines.hpp" 39#include "runtime/thread.inline.hpp" 40#include "services/threadService.hpp" 41#include "trace/tracing.hpp" 42#include "trace/traceMacros.hpp" 43#include "utilities/dtrace.hpp" 44#include "utilities/macros.hpp" 45#include "utilities/preserveException.hpp" 46 47#if defined(__GNUC__) && !defined(IA64) && !defined(PPC64) 48 // Need to inhibit inlining for older versions of GCC to avoid build-time failures 49 #define ATTR __attribute__((noinline)) 50#else 51 #define ATTR 52#endif 53 54 55#ifdef DTRACE_ENABLED 56 57// Only bother with this argument setup if dtrace is available 58// TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. 59 60 61#define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \ 62 char* bytes = NULL; \ 63 int len = 0; \ 64 jlong jtid = SharedRuntime::get_java_tid(thread); \ 65 Symbol* klassname = ((oop)obj)->klass()->name(); \ 66 if (klassname != NULL) { \ 67 bytes = (char*)klassname->bytes(); \ 68 len = klassname->utf8_length(); \ 69 } 70 71#define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 72 { \ 73 if (DTraceMonitorProbes) { \ 74 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 75 HOTSPOT_MONITOR_WAIT(jtid, \ 76 (monitor), bytes, len, (millis)); \ 77 } \ 78 } 79 80#define HOTSPOT_MONITOR_contended__enter HOTSPOT_MONITOR_CONTENDED_ENTER 81#define HOTSPOT_MONITOR_contended__entered HOTSPOT_MONITOR_CONTENDED_ENTERED 82#define HOTSPOT_MONITOR_contended__exit HOTSPOT_MONITOR_CONTENDED_EXIT 83#define HOTSPOT_MONITOR_notify HOTSPOT_MONITOR_NOTIFY 84#define HOTSPOT_MONITOR_notifyAll HOTSPOT_MONITOR_NOTIFYALL 85 86#define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 87 { \ 88 if (DTraceMonitorProbes) { \ 89 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 90 HOTSPOT_MONITOR_##probe(jtid, \ 91 (uintptr_t)(monitor), bytes, len); \ 92 } \ 93 } 94 95#else // ndef DTRACE_ENABLED 96 97#define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;} 98#define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;} 99 100#endif // ndef DTRACE_ENABLED 101 102// Tunables ... 103// The knob* variables are effectively final. Once set they should 104// never be modified hence. Consider using __read_mostly with GCC. 105 106int ObjectMonitor::Knob_Verbose = 0; 107int ObjectMonitor::Knob_SpinLimit = 5000; // derived by an external tool - 108static int Knob_LogSpins = 0; // enable jvmstat tally for spins 109static int Knob_HandOff = 0; 110static int Knob_ReportSettings = 0; 111 112static int Knob_SpinBase = 0; // Floor AKA SpinMin 113static int Knob_SpinBackOff = 0; // spin-loop backoff 114static int Knob_CASPenalty = -1; // Penalty for failed CAS 115static int Knob_OXPenalty = -1; // Penalty for observed _owner change 116static int Knob_SpinSetSucc = 1; // spinners set the _succ field 117static int Knob_SpinEarly = 1; 118static int Knob_SuccEnabled = 1; // futile wake throttling 119static int Knob_SuccRestrict = 0; // Limit successors + spinners to at-most-one 120static int Knob_MaxSpinners = -1; // Should be a function of # CPUs 121static int Knob_Bonus = 100; // spin success bonus 122static int Knob_BonusB = 100; // spin success bonus 123static int Knob_Penalty = 200; // spin failure penalty 124static int Knob_Poverty = 1000; 125static int Knob_SpinAfterFutile = 1; // Spin after returning from park() 126static int Knob_FixedSpin = 0; 127static int Knob_OState = 3; // Spinner checks thread state of _owner 128static int Knob_UsePause = 1; 129static int Knob_ExitPolicy = 0; 130static int Knob_PreSpin = 10; // 20-100 likely better 131static int Knob_ResetEvent = 0; 132static int BackOffMask = 0; 133 134static int Knob_FastHSSEC = 0; 135static int Knob_MoveNotifyee = 2; // notify() - disposition of notifyee 136static int Knob_QMode = 0; // EntryList-cxq policy - queue discipline 137static volatile int InitDone = 0; 138 139#define TrySpin TrySpin_VaryDuration 140 141// ----------------------------------------------------------------------------- 142// Theory of operations -- Monitors lists, thread residency, etc: 143// 144// * A thread acquires ownership of a monitor by successfully 145// CAS()ing the _owner field from null to non-null. 146// 147// * Invariant: A thread appears on at most one monitor list -- 148// cxq, EntryList or WaitSet -- at any one time. 149// 150// * Contending threads "push" themselves onto the cxq with CAS 151// and then spin/park. 152// 153// * After a contending thread eventually acquires the lock it must 154// dequeue itself from either the EntryList or the cxq. 155// 156// * The exiting thread identifies and unparks an "heir presumptive" 157// tentative successor thread on the EntryList. Critically, the 158// exiting thread doesn't unlink the successor thread from the EntryList. 159// After having been unparked, the wakee will recontend for ownership of 160// the monitor. The successor (wakee) will either acquire the lock or 161// re-park itself. 162// 163// Succession is provided for by a policy of competitive handoff. 164// The exiting thread does _not_ grant or pass ownership to the 165// successor thread. (This is also referred to as "handoff" succession"). 166// Instead the exiting thread releases ownership and possibly wakes 167// a successor, so the successor can (re)compete for ownership of the lock. 168// If the EntryList is empty but the cxq is populated the exiting 169// thread will drain the cxq into the EntryList. It does so by 170// by detaching the cxq (installing null with CAS) and folding 171// the threads from the cxq into the EntryList. The EntryList is 172// doubly linked, while the cxq is singly linked because of the 173// CAS-based "push" used to enqueue recently arrived threads (RATs). 174// 175// * Concurrency invariants: 176// 177// -- only the monitor owner may access or mutate the EntryList. 178// The mutex property of the monitor itself protects the EntryList 179// from concurrent interference. 180// -- Only the monitor owner may detach the cxq. 181// 182// * The monitor entry list operations avoid locks, but strictly speaking 183// they're not lock-free. Enter is lock-free, exit is not. 184// See http://j2se.east/~dice/PERSIST/040825-LockFreeQueues.html 185// 186// * The cxq can have multiple concurrent "pushers" but only one concurrent 187// detaching thread. This mechanism is immune from the ABA corruption. 188// More precisely, the CAS-based "push" onto cxq is ABA-oblivious. 189// 190// * Taken together, the cxq and the EntryList constitute or form a 191// single logical queue of threads stalled trying to acquire the lock. 192// We use two distinct lists to improve the odds of a constant-time 193// dequeue operation after acquisition (in the ::enter() epilogue) and 194// to reduce heat on the list ends. (c.f. Michael Scott's "2Q" algorithm). 195// A key desideratum is to minimize queue & monitor metadata manipulation 196// that occurs while holding the monitor lock -- that is, we want to 197// minimize monitor lock holds times. Note that even a small amount of 198// fixed spinning will greatly reduce the # of enqueue-dequeue operations 199// on EntryList|cxq. That is, spinning relieves contention on the "inner" 200// locks and monitor metadata. 201// 202// Cxq points to the the set of Recently Arrived Threads attempting entry. 203// Because we push threads onto _cxq with CAS, the RATs must take the form of 204// a singly-linked LIFO. We drain _cxq into EntryList at unlock-time when 205// the unlocking thread notices that EntryList is null but _cxq is != null. 206// 207// The EntryList is ordered by the prevailing queue discipline and 208// can be organized in any convenient fashion, such as a doubly-linked list or 209// a circular doubly-linked list. Critically, we want insert and delete operations 210// to operate in constant-time. If we need a priority queue then something akin 211// to Solaris' sleepq would work nicely. Viz., 212// http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c. 213// Queue discipline is enforced at ::exit() time, when the unlocking thread 214// drains the cxq into the EntryList, and orders or reorders the threads on the 215// EntryList accordingly. 216// 217// Barring "lock barging", this mechanism provides fair cyclic ordering, 218// somewhat similar to an elevator-scan. 219// 220// * The monitor synchronization subsystem avoids the use of native 221// synchronization primitives except for the narrow platform-specific 222// park-unpark abstraction. See the comments in os_solaris.cpp regarding 223// the semantics of park-unpark. Put another way, this monitor implementation 224// depends only on atomic operations and park-unpark. The monitor subsystem 225// manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the 226// underlying OS manages the READY<->RUN transitions. 227// 228// * Waiting threads reside on the WaitSet list -- wait() puts 229// the caller onto the WaitSet. 230// 231// * notify() or notifyAll() simply transfers threads from the WaitSet to 232// either the EntryList or cxq. Subsequent exit() operations will 233// unpark the notifyee. Unparking a notifee in notify() is inefficient - 234// it's likely the notifyee would simply impale itself on the lock held 235// by the notifier. 236// 237// * An interesting alternative is to encode cxq as (List,LockByte) where 238// the LockByte is 0 iff the monitor is owned. _owner is simply an auxiliary 239// variable, like _recursions, in the scheme. The threads or Events that form 240// the list would have to be aligned in 256-byte addresses. A thread would 241// try to acquire the lock or enqueue itself with CAS, but exiting threads 242// could use a 1-0 protocol and simply STB to set the LockByte to 0. 243// Note that is is *not* word-tearing, but it does presume that full-word 244// CAS operations are coherent with intermix with STB operations. That's true 245// on most common processors. 246// 247// * See also http://blogs.sun.com/dave 248 249 250// ----------------------------------------------------------------------------- 251// Enter support 252 253bool ObjectMonitor::try_enter(Thread* THREAD) { 254 if (THREAD != _owner) { 255 if (THREAD->is_lock_owned ((address)_owner)) { 256 assert(_recursions == 0, "internal state error"); 257 _owner = THREAD; 258 _recursions = 1; 259 OwnerIsThread = 1; 260 return true; 261 } 262 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 263 return false; 264 } 265 return true; 266 } else { 267 _recursions++; 268 return true; 269 } 270} 271 272void ATTR ObjectMonitor::enter(TRAPS) { 273 // The following code is ordered to check the most common cases first 274 // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors. 275 Thread * const Self = THREAD; 276 void * cur; 277 278 cur = Atomic::cmpxchg_ptr(Self, &_owner, NULL); 279 if (cur == NULL) { 280 // Either ASSERT _recursions == 0 or explicitly set _recursions = 0. 281 assert(_recursions == 0 , "invariant"); 282 assert(_owner == Self, "invariant"); 283 // CONSIDER: set or assert OwnerIsThread == 1 284 return; 285 } 286 287 if (cur == Self) { 288 // TODO-FIXME: check for integer overflow! BUGID 6557169. 289 _recursions++; 290 return; 291 } 292 293 if (Self->is_lock_owned ((address)cur)) { 294 assert(_recursions == 0, "internal state error"); 295 _recursions = 1; 296 // Commute owner from a thread-specific on-stack BasicLockObject address to 297 // a full-fledged "Thread *". 298 _owner = Self; 299 OwnerIsThread = 1; 300 return; 301 } 302 303 // We've encountered genuine contention. 304 assert(Self->_Stalled == 0, "invariant"); 305 Self->_Stalled = intptr_t(this); 306 307 // Try one round of spinning *before* enqueueing Self 308 // and before going through the awkward and expensive state 309 // transitions. The following spin is strictly optional ... 310 // Note that if we acquire the monitor from an initial spin 311 // we forgo posting JVMTI events and firing DTRACE probes. 312 if (Knob_SpinEarly && TrySpin (Self) > 0) { 313 assert(_owner == Self , "invariant"); 314 assert(_recursions == 0 , "invariant"); 315 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 316 Self->_Stalled = 0; 317 return; 318 } 319 320 assert(_owner != Self , "invariant"); 321 assert(_succ != Self , "invariant"); 322 assert(Self->is_Java_thread() , "invariant"); 323 JavaThread * jt = (JavaThread *) Self; 324 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 325 assert(jt->thread_state() != _thread_blocked , "invariant"); 326 assert(this->object() != NULL , "invariant"); 327 assert(_count >= 0, "invariant"); 328 329 // Prevent deflation at STW-time. See deflate_idle_monitors() and is_busy(). 330 // Ensure the object-monitor relationship remains stable while there's contention. 331 Atomic::inc_ptr(&_count); 332 333 EventJavaMonitorEnter event; 334 335 { // Change java thread status to indicate blocked on monitor enter. 336 JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this); 337 338 DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt); 339 if (JvmtiExport::should_post_monitor_contended_enter()) { 340 JvmtiExport::post_monitor_contended_enter(jt, this); 341 342 // The current thread does not yet own the monitor and does not 343 // yet appear on any queues that would get it made the successor. 344 // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event 345 // handler cannot accidentally consume an unpark() meant for the 346 // ParkEvent associated with this ObjectMonitor. 347 } 348 349 OSThreadContendState osts(Self->osthread()); 350 ThreadBlockInVM tbivm(jt); 351 352 Self->set_current_pending_monitor(this); 353 354 // TODO-FIXME: change the following for(;;) loop to straight-line code. 355 for (;;) { 356 jt->set_suspend_equivalent(); 357 // cleared by handle_special_suspend_equivalent_condition() 358 // or java_suspend_self() 359 360 EnterI(THREAD); 361 362 if (!ExitSuspendEquivalent(jt)) break; 363 364 // 365 // We have acquired the contended monitor, but while we were 366 // waiting another thread suspended us. We don't want to enter 367 // the monitor while suspended because that would surprise the 368 // thread that suspended us. 369 // 370 _recursions = 0; 371 _succ = NULL; 372 exit(false, Self); 373 374 jt->java_suspend_self(); 375 } 376 Self->set_current_pending_monitor(NULL); 377 378 // We cleared the pending monitor info since we've just gotten past 379 // the enter-check-for-suspend dance and we now own the monitor free 380 // and clear, i.e., it is no longer pending. The ThreadBlockInVM 381 // destructor can go to a safepoint at the end of this block. If we 382 // do a thread dump during that safepoint, then this thread will show 383 // as having "-locked" the monitor, but the OS and java.lang.Thread 384 // states will still report that the thread is blocked trying to 385 // acquire it. 386 } 387 388 Atomic::dec_ptr(&_count); 389 assert(_count >= 0, "invariant"); 390 Self->_Stalled = 0; 391 392 // Must either set _recursions = 0 or ASSERT _recursions == 0. 393 assert(_recursions == 0 , "invariant"); 394 assert(_owner == Self , "invariant"); 395 assert(_succ != Self , "invariant"); 396 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 397 398 // The thread -- now the owner -- is back in vm mode. 399 // Report the glorious news via TI,DTrace and jvmstat. 400 // The probe effect is non-trivial. All the reportage occurs 401 // while we hold the monitor, increasing the length of the critical 402 // section. Amdahl's parallel speedup law comes vividly into play. 403 // 404 // Another option might be to aggregate the events (thread local or 405 // per-monitor aggregation) and defer reporting until a more opportune 406 // time -- such as next time some thread encounters contention but has 407 // yet to acquire the lock. While spinning that thread could 408 // spinning we could increment JVMStat counters, etc. 409 410 DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt); 411 if (JvmtiExport::should_post_monitor_contended_entered()) { 412 JvmtiExport::post_monitor_contended_entered(jt, this); 413 414 // The current thread already owns the monitor and is not going to 415 // call park() for the remainder of the monitor enter protocol. So 416 // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED 417 // event handler consumed an unpark() issued by the thread that 418 // just exited the monitor. 419 } 420 421 if (event.should_commit()) { 422 event.set_klass(((oop)this->object())->klass()); 423 event.set_previousOwner((TYPE_JAVALANGTHREAD)_previous_owner_tid); 424 event.set_address((TYPE_ADDRESS)(uintptr_t)(this->object_addr())); 425 event.commit(); 426 } 427 428 if (ObjectMonitor::_sync_ContendedLockAttempts != NULL) { 429 ObjectMonitor::_sync_ContendedLockAttempts->inc(); 430 } 431} 432 433 434// Caveat: TryLock() is not necessarily serializing if it returns failure. 435// Callers must compensate as needed. 436 437int ObjectMonitor::TryLock (Thread * Self) { 438 for (;;) { 439 void * own = _owner; 440 if (own != NULL) return 0; 441 if (Atomic::cmpxchg_ptr (Self, &_owner, NULL) == NULL) { 442 // Either guarantee _recursions == 0 or set _recursions = 0. 443 assert(_recursions == 0, "invariant"); 444 assert(_owner == Self, "invariant"); 445 // CONSIDER: set or assert that OwnerIsThread == 1 446 return 1; 447 } 448 // The lock had been free momentarily, but we lost the race to the lock. 449 // Interference -- the CAS failed. 450 // We can either return -1 or retry. 451 // Retry doesn't make as much sense because the lock was just acquired. 452 if (true) return -1; 453 } 454} 455 456void ATTR ObjectMonitor::EnterI (TRAPS) { 457 Thread * Self = THREAD; 458 assert(Self->is_Java_thread(), "invariant"); 459 assert(((JavaThread *) Self)->thread_state() == _thread_blocked , "invariant"); 460 461 // Try the lock - TATAS 462 if (TryLock (Self) > 0) { 463 assert(_succ != Self , "invariant"); 464 assert(_owner == Self , "invariant"); 465 assert(_Responsible != Self , "invariant"); 466 return; 467 } 468 469 DeferredInitialize(); 470 471 // We try one round of spinning *before* enqueueing Self. 472 // 473 // If the _owner is ready but OFFPROC we could use a YieldTo() 474 // operation to donate the remainder of this thread's quantum 475 // to the owner. This has subtle but beneficial affinity 476 // effects. 477 478 if (TrySpin (Self) > 0) { 479 assert(_owner == Self , "invariant"); 480 assert(_succ != Self , "invariant"); 481 assert(_Responsible != Self , "invariant"); 482 return; 483 } 484 485 // The Spin failed -- Enqueue and park the thread ... 486 assert(_succ != Self , "invariant"); 487 assert(_owner != Self , "invariant"); 488 assert(_Responsible != Self , "invariant"); 489 490 // Enqueue "Self" on ObjectMonitor's _cxq. 491 // 492 // Node acts as a proxy for Self. 493 // As an aside, if were to ever rewrite the synchronization code mostly 494 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class 495 // Java objects. This would avoid awkward lifecycle and liveness issues, 496 // as well as eliminate a subset of ABA issues. 497 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events. 498 // 499 500 ObjectWaiter node(Self); 501 Self->_ParkEvent->reset(); 502 node._prev = (ObjectWaiter *) 0xBAD; 503 node.TState = ObjectWaiter::TS_CXQ; 504 505 // Push "Self" onto the front of the _cxq. 506 // Once on cxq/EntryList, Self stays on-queue until it acquires the lock. 507 // Note that spinning tends to reduce the rate at which threads 508 // enqueue and dequeue on EntryList|cxq. 509 ObjectWaiter * nxt; 510 for (;;) { 511 node._next = nxt = _cxq; 512 if (Atomic::cmpxchg_ptr(&node, &_cxq, nxt) == nxt) break; 513 514 // Interference - the CAS failed because _cxq changed. Just retry. 515 // As an optional optimization we retry the lock. 516 if (TryLock (Self) > 0) { 517 assert(_succ != Self , "invariant"); 518 assert(_owner == Self , "invariant"); 519 assert(_Responsible != Self , "invariant"); 520 return; 521 } 522 } 523 524 // Check for cxq|EntryList edge transition to non-null. This indicates 525 // the onset of contention. While contention persists exiting threads 526 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit 527 // operations revert to the faster 1-0 mode. This enter operation may interleave 528 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we 529 // arrange for one of the contending thread to use a timed park() operations 530 // to detect and recover from the race. (Stranding is form of progress failure 531 // where the monitor is unlocked but all the contending threads remain parked). 532 // That is, at least one of the contended threads will periodically poll _owner. 533 // One of the contending threads will become the designated "Responsible" thread. 534 // The Responsible thread uses a timed park instead of a normal indefinite park 535 // operation -- it periodically wakes and checks for and recovers from potential 536 // strandings admitted by 1-0 exit operations. We need at most one Responsible 537 // thread per-monitor at any given moment. Only threads on cxq|EntryList may 538 // be responsible for a monitor. 539 // 540 // Currently, one of the contended threads takes on the added role of "Responsible". 541 // A viable alternative would be to use a dedicated "stranding checker" thread 542 // that periodically iterated over all the threads (or active monitors) and unparked 543 // successors where there was risk of stranding. This would help eliminate the 544 // timer scalability issues we see on some platforms as we'd only have one thread 545 // -- the checker -- parked on a timer. 546 547 if ((SyncFlags & 16) == 0 && nxt == NULL && _EntryList == NULL) { 548 // Try to assume the role of responsible thread for the monitor. 549 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=Self } 550 Atomic::cmpxchg_ptr(Self, &_Responsible, NULL); 551 } 552 553 // The lock have been released while this thread was occupied queueing 554 // itself onto _cxq. To close the race and avoid "stranding" and 555 // progress-liveness failure we must resample-retry _owner before parking. 556 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner. 557 // In this case the ST-MEMBAR is accomplished with CAS(). 558 // 559 // TODO: Defer all thread state transitions until park-time. 560 // Since state transitions are heavy and inefficient we'd like 561 // to defer the state transitions until absolutely necessary, 562 // and in doing so avoid some transitions ... 563 564 TEVENT(Inflated enter - Contention); 565 int nWakeups = 0; 566 int RecheckInterval = 1; 567 568 for (;;) { 569 570 if (TryLock(Self) > 0) break; 571 assert(_owner != Self, "invariant"); 572 573 if ((SyncFlags & 2) && _Responsible == NULL) { 574 Atomic::cmpxchg_ptr(Self, &_Responsible, NULL); 575 } 576 577 // park self 578 if (_Responsible == Self || (SyncFlags & 1)) { 579 TEVENT(Inflated enter - park TIMED); 580 Self->_ParkEvent->park((jlong) RecheckInterval); 581 // Increase the RecheckInterval, but clamp the value. 582 RecheckInterval *= 8; 583 if (RecheckInterval > 1000) RecheckInterval = 1000; 584 } else { 585 TEVENT(Inflated enter - park UNTIMED); 586 Self->_ParkEvent->park(); 587 } 588 589 if (TryLock(Self) > 0) break; 590 591 // The lock is still contested. 592 // Keep a tally of the # of futile wakeups. 593 // Note that the counter is not protected by a lock or updated by atomics. 594 // That is by design - we trade "lossy" counters which are exposed to 595 // races during updates for a lower probe effect. 596 TEVENT(Inflated enter - Futile wakeup); 597 if (ObjectMonitor::_sync_FutileWakeups != NULL) { 598 ObjectMonitor::_sync_FutileWakeups->inc(); 599 } 600 ++nWakeups; 601 602 // Assuming this is not a spurious wakeup we'll normally find _succ == Self. 603 // We can defer clearing _succ until after the spin completes 604 // TrySpin() must tolerate being called with _succ == Self. 605 // Try yet another round of adaptive spinning. 606 if ((Knob_SpinAfterFutile & 1) && TrySpin(Self) > 0) break; 607 608 // We can find that we were unpark()ed and redesignated _succ while 609 // we were spinning. That's harmless. If we iterate and call park(), 610 // park() will consume the event and return immediately and we'll 611 // just spin again. This pattern can repeat, leaving _succ to simply 612 // spin on a CPU. Enable Knob_ResetEvent to clear pending unparks(). 613 // Alternately, we can sample fired() here, and if set, forgo spinning 614 // in the next iteration. 615 616 if ((Knob_ResetEvent & 1) && Self->_ParkEvent->fired()) { 617 Self->_ParkEvent->reset(); 618 OrderAccess::fence(); 619 } 620 if (_succ == Self) _succ = NULL; 621 622 // Invariant: after clearing _succ a thread *must* retry _owner before parking. 623 OrderAccess::fence(); 624 } 625 626 // Egress : 627 // Self has acquired the lock -- Unlink Self from the cxq or EntryList. 628 // Normally we'll find Self on the EntryList . 629 // From the perspective of the lock owner (this thread), the 630 // EntryList is stable and cxq is prepend-only. 631 // The head of cxq is volatile but the interior is stable. 632 // In addition, Self.TState is stable. 633 634 assert(_owner == Self , "invariant"); 635 assert(object() != NULL , "invariant"); 636 // I'd like to write: 637 // guarantee (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; 638 // but as we're at a safepoint that's not safe. 639 640 UnlinkAfterAcquire(Self, &node); 641 if (_succ == Self) _succ = NULL; 642 643 assert(_succ != Self, "invariant"); 644 if (_Responsible == Self) { 645 _Responsible = NULL; 646 OrderAccess::fence(); // Dekker pivot-point 647 648 // We may leave threads on cxq|EntryList without a designated 649 // "Responsible" thread. This is benign. When this thread subsequently 650 // exits the monitor it can "see" such preexisting "old" threads -- 651 // threads that arrived on the cxq|EntryList before the fence, above -- 652 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads 653 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible 654 // non-null and elect a new "Responsible" timer thread. 655 // 656 // This thread executes: 657 // ST Responsible=null; MEMBAR (in enter epilogue - here) 658 // LD cxq|EntryList (in subsequent exit) 659 // 660 // Entering threads in the slow/contended path execute: 661 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog) 662 // The (ST cxq; MEMBAR) is accomplished with CAS(). 663 // 664 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent 665 // exit operation from floating above the ST Responsible=null. 666 } 667 668 // We've acquired ownership with CAS(). 669 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics. 670 // But since the CAS() this thread may have also stored into _succ, 671 // EntryList, cxq or Responsible. These meta-data updates must be 672 // visible __before this thread subsequently drops the lock. 673 // Consider what could occur if we didn't enforce this constraint -- 674 // STs to monitor meta-data and user-data could reorder with (become 675 // visible after) the ST in exit that drops ownership of the lock. 676 // Some other thread could then acquire the lock, but observe inconsistent 677 // or old monitor meta-data and heap data. That violates the JMM. 678 // To that end, the 1-0 exit() operation must have at least STST|LDST 679 // "release" barrier semantics. Specifically, there must be at least a 680 // STST|LDST barrier in exit() before the ST of null into _owner that drops 681 // the lock. The barrier ensures that changes to monitor meta-data and data 682 // protected by the lock will be visible before we release the lock, and 683 // therefore before some other thread (CPU) has a chance to acquire the lock. 684 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html. 685 // 686 // Critically, any prior STs to _succ or EntryList must be visible before 687 // the ST of null into _owner in the *subsequent* (following) corresponding 688 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily 689 // execute a serializing instruction. 690 691 if (SyncFlags & 8) { 692 OrderAccess::fence(); 693 } 694 return; 695} 696 697// ReenterI() is a specialized inline form of the latter half of the 698// contended slow-path from EnterI(). We use ReenterI() only for 699// monitor reentry in wait(). 700// 701// In the future we should reconcile EnterI() and ReenterI(), adding 702// Knob_Reset and Knob_SpinAfterFutile support and restructuring the 703// loop accordingly. 704 705void ATTR ObjectMonitor::ReenterI (Thread * Self, ObjectWaiter * SelfNode) { 706 assert(Self != NULL , "invariant"); 707 assert(SelfNode != NULL , "invariant"); 708 assert(SelfNode->_thread == Self , "invariant"); 709 assert(_waiters > 0 , "invariant"); 710 assert(((oop)(object()))->mark() == markOopDesc::encode(this) , "invariant"); 711 assert(((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant"); 712 JavaThread * jt = (JavaThread *) Self; 713 714 int nWakeups = 0; 715 for (;;) { 716 ObjectWaiter::TStates v = SelfNode->TState; 717 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 718 assert(_owner != Self, "invariant"); 719 720 if (TryLock(Self) > 0) break; 721 if (TrySpin(Self) > 0) break; 722 723 TEVENT(Wait Reentry - parking); 724 725 // State transition wrappers around park() ... 726 // ReenterI() wisely defers state transitions until 727 // it's clear we must park the thread. 728 { 729 OSThreadContendState osts(Self->osthread()); 730 ThreadBlockInVM tbivm(jt); 731 732 // cleared by handle_special_suspend_equivalent_condition() 733 // or java_suspend_self() 734 jt->set_suspend_equivalent(); 735 if (SyncFlags & 1) { 736 Self->_ParkEvent->park((jlong)1000); 737 } else { 738 Self->_ParkEvent->park(); 739 } 740 741 // were we externally suspended while we were waiting? 742 for (;;) { 743 if (!ExitSuspendEquivalent(jt)) break; 744 if (_succ == Self) { _succ = NULL; OrderAccess::fence(); } 745 jt->java_suspend_self(); 746 jt->set_suspend_equivalent(); 747 } 748 } 749 750 // Try again, but just so we distinguish between futile wakeups and 751 // successful wakeups. The following test isn't algorithmically 752 // necessary, but it helps us maintain sensible statistics. 753 if (TryLock(Self) > 0) break; 754 755 // The lock is still contested. 756 // Keep a tally of the # of futile wakeups. 757 // Note that the counter is not protected by a lock or updated by atomics. 758 // That is by design - we trade "lossy" counters which are exposed to 759 // races during updates for a lower probe effect. 760 TEVENT(Wait Reentry - futile wakeup); 761 ++nWakeups; 762 763 // Assuming this is not a spurious wakeup we'll normally 764 // find that _succ == Self. 765 if (_succ == Self) _succ = NULL; 766 767 // Invariant: after clearing _succ a contending thread 768 // *must* retry _owner before parking. 769 OrderAccess::fence(); 770 771 if (ObjectMonitor::_sync_FutileWakeups != NULL) { 772 ObjectMonitor::_sync_FutileWakeups->inc(); 773 } 774 } 775 776 // Self has acquired the lock -- Unlink Self from the cxq or EntryList . 777 // Normally we'll find Self on the EntryList. 778 // Unlinking from the EntryList is constant-time and atomic-free. 779 // From the perspective of the lock owner (this thread), the 780 // EntryList is stable and cxq is prepend-only. 781 // The head of cxq is volatile but the interior is stable. 782 // In addition, Self.TState is stable. 783 784 assert(_owner == Self, "invariant"); 785 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 786 UnlinkAfterAcquire(Self, SelfNode); 787 if (_succ == Self) _succ = NULL; 788 assert(_succ != Self, "invariant"); 789 SelfNode->TState = ObjectWaiter::TS_RUN; 790 OrderAccess::fence(); // see comments at the end of EnterI() 791} 792 793// after the thread acquires the lock in ::enter(). Equally, we could defer 794// unlinking the thread until ::exit()-time. 795 796void ObjectMonitor::UnlinkAfterAcquire (Thread * Self, ObjectWaiter * SelfNode) 797{ 798 assert(_owner == Self, "invariant"); 799 assert(SelfNode->_thread == Self, "invariant"); 800 801 if (SelfNode->TState == ObjectWaiter::TS_ENTER) { 802 // Normal case: remove Self from the DLL EntryList . 803 // This is a constant-time operation. 804 ObjectWaiter * nxt = SelfNode->_next; 805 ObjectWaiter * prv = SelfNode->_prev; 806 if (nxt != NULL) nxt->_prev = prv; 807 if (prv != NULL) prv->_next = nxt; 808 if (SelfNode == _EntryList) _EntryList = nxt; 809 assert(nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant"); 810 assert(prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant"); 811 TEVENT(Unlink from EntryList); 812 } else { 813 guarantee(SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant"); 814 // Inopportune interleaving -- Self is still on the cxq. 815 // This usually means the enqueue of self raced an exiting thread. 816 // Normally we'll find Self near the front of the cxq, so 817 // dequeueing is typically fast. If needbe we can accelerate 818 // this with some MCS/CHL-like bidirectional list hints and advisory 819 // back-links so dequeueing from the interior will normally operate 820 // in constant-time. 821 // Dequeue Self from either the head (with CAS) or from the interior 822 // with a linear-time scan and normal non-atomic memory operations. 823 // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList 824 // and then unlink Self from EntryList. We have to drain eventually, 825 // so it might as well be now. 826 827 ObjectWaiter * v = _cxq; 828 assert(v != NULL, "invariant"); 829 if (v != SelfNode || Atomic::cmpxchg_ptr (SelfNode->_next, &_cxq, v) != v) { 830 // The CAS above can fail from interference IFF a "RAT" arrived. 831 // In that case Self must be in the interior and can no longer be 832 // at the head of cxq. 833 if (v == SelfNode) { 834 assert(_cxq != v, "invariant"); 835 v = _cxq; // CAS above failed - start scan at head of list 836 } 837 ObjectWaiter * p; 838 ObjectWaiter * q = NULL; 839 for (p = v; p != NULL && p != SelfNode; p = p->_next) { 840 q = p; 841 assert(p->TState == ObjectWaiter::TS_CXQ, "invariant"); 842 } 843 assert(v != SelfNode, "invariant"); 844 assert(p == SelfNode, "Node not found on cxq"); 845 assert(p != _cxq, "invariant"); 846 assert(q != NULL, "invariant"); 847 assert(q->_next == p, "invariant"); 848 q->_next = p->_next; 849 } 850 TEVENT(Unlink from cxq); 851 } 852 853 // Diagnostic hygiene ... 854 SelfNode->_prev = (ObjectWaiter *) 0xBAD; 855 SelfNode->_next = (ObjectWaiter *) 0xBAD; 856 SelfNode->TState = ObjectWaiter::TS_RUN; 857} 858 859// ----------------------------------------------------------------------------- 860// Exit support 861// 862// exit() 863// ~~~~~~ 864// Note that the collector can't reclaim the objectMonitor or deflate 865// the object out from underneath the thread calling ::exit() as the 866// thread calling ::exit() never transitions to a stable state. 867// This inhibits GC, which in turn inhibits asynchronous (and 868// inopportune) reclamation of "this". 869// 870// We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ; 871// There's one exception to the claim above, however. EnterI() can call 872// exit() to drop a lock if the acquirer has been externally suspended. 873// In that case exit() is called with _thread_state as _thread_blocked, 874// but the monitor's _count field is > 0, which inhibits reclamation. 875// 876// 1-0 exit 877// ~~~~~~~~ 878// ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of 879// the fast-path operators have been optimized so the common ::exit() 880// operation is 1-0. See i486.ad fast_unlock(), for instance. 881// The code emitted by fast_unlock() elides the usual MEMBAR. This 882// greatly improves latency -- MEMBAR and CAS having considerable local 883// latency on modern processors -- but at the cost of "stranding". Absent the 884// MEMBAR, a thread in fast_unlock() can race a thread in the slow 885// ::enter() path, resulting in the entering thread being stranding 886// and a progress-liveness failure. Stranding is extremely rare. 887// We use timers (timed park operations) & periodic polling to detect 888// and recover from stranding. Potentially stranded threads periodically 889// wake up and poll the lock. See the usage of the _Responsible variable. 890// 891// The CAS() in enter provides for safety and exclusion, while the CAS or 892// MEMBAR in exit provides for progress and avoids stranding. 1-0 locking 893// eliminates the CAS/MEMBAR from the exist path, but it admits stranding. 894// We detect and recover from stranding with timers. 895// 896// If a thread transiently strands it'll park until (a) another 897// thread acquires the lock and then drops the lock, at which time the 898// exiting thread will notice and unpark the stranded thread, or, (b) 899// the timer expires. If the lock is high traffic then the stranding latency 900// will be low due to (a). If the lock is low traffic then the odds of 901// stranding are lower, although the worst-case stranding latency 902// is longer. Critically, we don't want to put excessive load in the 903// platform's timer subsystem. We want to minimize both the timer injection 904// rate (timers created/sec) as well as the number of timers active at 905// any one time. (more precisely, we want to minimize timer-seconds, which is 906// the integral of the # of active timers at any instant over time). 907// Both impinge on OS scalability. Given that, at most one thread parked on 908// a monitor will use a timer. 909 910void ATTR ObjectMonitor::exit(bool not_suspended, TRAPS) { 911 Thread * Self = THREAD; 912 if (THREAD != _owner) { 913 if (THREAD->is_lock_owned((address) _owner)) { 914 // Transmute _owner from a BasicLock pointer to a Thread address. 915 // We don't need to hold _mutex for this transition. 916 // Non-null to Non-null is safe as long as all readers can 917 // tolerate either flavor. 918 assert(_recursions == 0, "invariant"); 919 _owner = THREAD; 920 _recursions = 0; 921 OwnerIsThread = 1; 922 } else { 923 // NOTE: we need to handle unbalanced monitor enter/exit 924 // in native code by throwing an exception. 925 // TODO: Throw an IllegalMonitorStateException ? 926 TEVENT(Exit - Throw IMSX); 927 assert(false, "Non-balanced monitor enter/exit!"); 928 if (false) { 929 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); 930 } 931 return; 932 } 933 } 934 935 if (_recursions != 0) { 936 _recursions--; // this is simple recursive enter 937 TEVENT(Inflated exit - recursive); 938 return; 939 } 940 941 // Invariant: after setting Responsible=null an thread must execute 942 // a MEMBAR or other serializing instruction before fetching EntryList|cxq. 943 if ((SyncFlags & 4) == 0) { 944 _Responsible = NULL; 945 } 946 947#if INCLUDE_TRACE 948 // get the owner's thread id for the MonitorEnter event 949 // if it is enabled and the thread isn't suspended 950 if (not_suspended && Tracing::is_event_enabled(TraceJavaMonitorEnterEvent)) { 951 _previous_owner_tid = SharedRuntime::get_java_tid(Self); 952 } 953#endif 954 955 for (;;) { 956 assert(THREAD == _owner, "invariant"); 957 958 959 if (Knob_ExitPolicy == 0) { 960 // release semantics: prior loads and stores from within the critical section 961 // must not float (reorder) past the following store that drops the lock. 962 // On SPARC that requires MEMBAR #loadstore|#storestore. 963 // But of course in TSO #loadstore|#storestore is not required. 964 // I'd like to write one of the following: 965 // A. OrderAccess::release() ; _owner = NULL 966 // B. OrderAccess::loadstore(); OrderAccess::storestore(); _owner = NULL; 967 // Unfortunately OrderAccess::release() and OrderAccess::loadstore() both 968 // store into a _dummy variable. That store is not needed, but can result 969 // in massive wasteful coherency traffic on classic SMP systems. 970 // Instead, I use release_store(), which is implemented as just a simple 971 // ST on x64, x86 and SPARC. 972 OrderAccess::release_store_ptr(&_owner, NULL); // drop the lock 973 OrderAccess::storeload(); // See if we need to wake a successor 974 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 975 TEVENT(Inflated exit - simple egress); 976 return; 977 } 978 TEVENT(Inflated exit - complex egress); 979 980 // Normally the exiting thread is responsible for ensuring succession, 981 // but if other successors are ready or other entering threads are spinning 982 // then this thread can simply store NULL into _owner and exit without 983 // waking a successor. The existence of spinners or ready successors 984 // guarantees proper succession (liveness). Responsibility passes to the 985 // ready or running successors. The exiting thread delegates the duty. 986 // More precisely, if a successor already exists this thread is absolved 987 // of the responsibility of waking (unparking) one. 988 // 989 // The _succ variable is critical to reducing futile wakeup frequency. 990 // _succ identifies the "heir presumptive" thread that has been made 991 // ready (unparked) but that has not yet run. We need only one such 992 // successor thread to guarantee progress. 993 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf 994 // section 3.3 "Futile Wakeup Throttling" for details. 995 // 996 // Note that spinners in Enter() also set _succ non-null. 997 // In the current implementation spinners opportunistically set 998 // _succ so that exiting threads might avoid waking a successor. 999 // Another less appealing alternative would be for the exiting thread 1000 // to drop the lock and then spin briefly to see if a spinner managed 1001 // to acquire the lock. If so, the exiting thread could exit 1002 // immediately without waking a successor, otherwise the exiting 1003 // thread would need to dequeue and wake a successor. 1004 // (Note that we'd need to make the post-drop spin short, but no 1005 // shorter than the worst-case round-trip cache-line migration time. 1006 // The dropped lock needs to become visible to the spinner, and then 1007 // the acquisition of the lock by the spinner must become visible to 1008 // the exiting thread). 1009 // 1010 1011 // It appears that an heir-presumptive (successor) must be made ready. 1012 // Only the current lock owner can manipulate the EntryList or 1013 // drain _cxq, so we need to reacquire the lock. If we fail 1014 // to reacquire the lock the responsibility for ensuring succession 1015 // falls to the new owner. 1016 // 1017 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 1018 return; 1019 } 1020 TEVENT(Exit - Reacquired); 1021 } else { 1022 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { 1023 OrderAccess::release_store_ptr(&_owner, NULL); // drop the lock 1024 OrderAccess::storeload(); 1025 // Ratify the previously observed values. 1026 if (_cxq == NULL || _succ != NULL) { 1027 TEVENT(Inflated exit - simple egress); 1028 return; 1029 } 1030 1031 // inopportune interleaving -- the exiting thread (this thread) 1032 // in the fast-exit path raced an entering thread in the slow-enter 1033 // path. 1034 // We have two choices: 1035 // A. Try to reacquire the lock. 1036 // If the CAS() fails return immediately, otherwise 1037 // we either restart/rerun the exit operation, or simply 1038 // fall-through into the code below which wakes a successor. 1039 // B. If the elements forming the EntryList|cxq are TSM 1040 // we could simply unpark() the lead thread and return 1041 // without having set _succ. 1042 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { 1043 TEVENT(Inflated exit - reacquired succeeded); 1044 return; 1045 } 1046 TEVENT(Inflated exit - reacquired failed); 1047 } else { 1048 TEVENT(Inflated exit - complex egress); 1049 } 1050 } 1051 1052 guarantee(_owner == THREAD, "invariant"); 1053 1054 ObjectWaiter * w = NULL; 1055 int QMode = Knob_QMode; 1056 1057 if (QMode == 2 && _cxq != NULL) { 1058 // QMode == 2 : cxq has precedence over EntryList. 1059 // Try to directly wake a successor from the cxq. 1060 // If successful, the successor will need to unlink itself from cxq. 1061 w = _cxq; 1062 assert(w != NULL, "invariant"); 1063 assert(w->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1064 ExitEpilog(Self, w); 1065 return; 1066 } 1067 1068 if (QMode == 3 && _cxq != NULL) { 1069 // Aggressively drain cxq into EntryList at the first opportunity. 1070 // This policy ensure that recently-run threads live at the head of EntryList. 1071 // Drain _cxq into EntryList - bulk transfer. 1072 // First, detach _cxq. 1073 // The following loop is tantamount to: w = swap (&cxq, NULL) 1074 w = _cxq; 1075 for (;;) { 1076 assert(w != NULL, "Invariant"); 1077 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr(NULL, &_cxq, w); 1078 if (u == w) break; 1079 w = u; 1080 } 1081 assert(w != NULL , "invariant"); 1082 1083 ObjectWaiter * q = NULL; 1084 ObjectWaiter * p; 1085 for (p = w; p != NULL; p = p->_next) { 1086 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1087 p->TState = ObjectWaiter::TS_ENTER; 1088 p->_prev = q; 1089 q = p; 1090 } 1091 1092 // Append the RATs to the EntryList 1093 // TODO: organize EntryList as a CDLL so we can locate the tail in constant-time. 1094 ObjectWaiter * Tail; 1095 for (Tail = _EntryList; Tail != NULL && Tail->_next != NULL; Tail = Tail->_next); 1096 if (Tail == NULL) { 1097 _EntryList = w; 1098 } else { 1099 Tail->_next = w; 1100 w->_prev = Tail; 1101 } 1102 1103 // Fall thru into code that tries to wake a successor from EntryList 1104 } 1105 1106 if (QMode == 4 && _cxq != NULL) { 1107 // Aggressively drain cxq into EntryList at the first opportunity. 1108 // This policy ensure that recently-run threads live at the head of EntryList. 1109 1110 // Drain _cxq into EntryList - bulk transfer. 1111 // First, detach _cxq. 1112 // The following loop is tantamount to: w = swap (&cxq, NULL) 1113 w = _cxq; 1114 for (;;) { 1115 assert(w != NULL, "Invariant"); 1116 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr(NULL, &_cxq, w); 1117 if (u == w) break; 1118 w = u; 1119 } 1120 assert(w != NULL , "invariant"); 1121 1122 ObjectWaiter * q = NULL; 1123 ObjectWaiter * p; 1124 for (p = w; p != NULL; p = p->_next) { 1125 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1126 p->TState = ObjectWaiter::TS_ENTER; 1127 p->_prev = q; 1128 q = p; 1129 } 1130 1131 // Prepend the RATs to the EntryList 1132 if (_EntryList != NULL) { 1133 q->_next = _EntryList; 1134 _EntryList->_prev = q; 1135 } 1136 _EntryList = w; 1137 1138 // Fall thru into code that tries to wake a successor from EntryList 1139 } 1140 1141 w = _EntryList; 1142 if (w != NULL) { 1143 // I'd like to write: guarantee (w->_thread != Self). 1144 // But in practice an exiting thread may find itself on the EntryList. 1145 // Lets say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and 1146 // then calls exit(). Exit release the lock by setting O._owner to NULL. 1147 // Lets say T1 then stalls. T2 acquires O and calls O.notify(). The 1148 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then 1149 // release the lock "O". T2 resumes immediately after the ST of null into 1150 // _owner, above. T2 notices that the EntryList is populated, so it 1151 // reacquires the lock and then finds itself on the EntryList. 1152 // Given all that, we have to tolerate the circumstance where "w" is 1153 // associated with Self. 1154 assert(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1155 ExitEpilog(Self, w); 1156 return; 1157 } 1158 1159 // If we find that both _cxq and EntryList are null then just 1160 // re-run the exit protocol from the top. 1161 w = _cxq; 1162 if (w == NULL) continue; 1163 1164 // Drain _cxq into EntryList - bulk transfer. 1165 // First, detach _cxq. 1166 // The following loop is tantamount to: w = swap (&cxq, NULL) 1167 for (;;) { 1168 assert(w != NULL, "Invariant"); 1169 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr(NULL, &_cxq, w); 1170 if (u == w) break; 1171 w = u; 1172 } 1173 TEVENT(Inflated exit - drain cxq into EntryList); 1174 1175 assert(w != NULL , "invariant"); 1176 assert(_EntryList == NULL , "invariant"); 1177 1178 // Convert the LIFO SLL anchored by _cxq into a DLL. 1179 // The list reorganization step operates in O(LENGTH(w)) time. 1180 // It's critical that this step operate quickly as 1181 // "Self" still holds the outer-lock, restricting parallelism 1182 // and effectively lengthening the critical section. 1183 // Invariant: s chases t chases u. 1184 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so 1185 // we have faster access to the tail. 1186 1187 if (QMode == 1) { 1188 // QMode == 1 : drain cxq to EntryList, reversing order 1189 // We also reverse the order of the list. 1190 ObjectWaiter * s = NULL; 1191 ObjectWaiter * t = w; 1192 ObjectWaiter * u = NULL; 1193 while (t != NULL) { 1194 guarantee(t->TState == ObjectWaiter::TS_CXQ, "invariant"); 1195 t->TState = ObjectWaiter::TS_ENTER; 1196 u = t->_next; 1197 t->_prev = u; 1198 t->_next = s; 1199 s = t; 1200 t = u; 1201 } 1202 _EntryList = s; 1203 assert(s != NULL, "invariant"); 1204 } else { 1205 // QMode == 0 or QMode == 2 1206 _EntryList = w; 1207 ObjectWaiter * q = NULL; 1208 ObjectWaiter * p; 1209 for (p = w; p != NULL; p = p->_next) { 1210 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant"); 1211 p->TState = ObjectWaiter::TS_ENTER; 1212 p->_prev = q; 1213 q = p; 1214 } 1215 } 1216 1217 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL 1218 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog(). 1219 1220 // See if we can abdicate to a spinner instead of waking a thread. 1221 // A primary goal of the implementation is to reduce the 1222 // context-switch rate. 1223 if (_succ != NULL) continue; 1224 1225 w = _EntryList; 1226 if (w != NULL) { 1227 guarantee(w->TState == ObjectWaiter::TS_ENTER, "invariant"); 1228 ExitEpilog(Self, w); 1229 return; 1230 } 1231 } 1232} 1233 1234// ExitSuspendEquivalent: 1235// A faster alternate to handle_special_suspend_equivalent_condition() 1236// 1237// handle_special_suspend_equivalent_condition() unconditionally 1238// acquires the SR_lock. On some platforms uncontended MutexLocker() 1239// operations have high latency. Note that in ::enter() we call HSSEC 1240// while holding the monitor, so we effectively lengthen the critical sections. 1241// 1242// There are a number of possible solutions: 1243// 1244// A. To ameliorate the problem we might also defer state transitions 1245// to as late as possible -- just prior to parking. 1246// Given that, we'd call HSSEC after having returned from park(), 1247// but before attempting to acquire the monitor. This is only a 1248// partial solution. It avoids calling HSSEC while holding the 1249// monitor (good), but it still increases successor reacquisition latency -- 1250// the interval between unparking a successor and the time the successor 1251// resumes and retries the lock. See ReenterI(), which defers state transitions. 1252// If we use this technique we can also avoid EnterI()-exit() loop 1253// in ::enter() where we iteratively drop the lock and then attempt 1254// to reacquire it after suspending. 1255// 1256// B. In the future we might fold all the suspend bits into a 1257// composite per-thread suspend flag and then update it with CAS(). 1258// Alternately, a Dekker-like mechanism with multiple variables 1259// would suffice: 1260// ST Self->_suspend_equivalent = false 1261// MEMBAR 1262// LD Self_>_suspend_flags 1263// 1264 1265 1266bool ObjectMonitor::ExitSuspendEquivalent (JavaThread * jSelf) { 1267 int Mode = Knob_FastHSSEC; 1268 if (Mode && !jSelf->is_external_suspend()) { 1269 assert(jSelf->is_suspend_equivalent(), "invariant"); 1270 jSelf->clear_suspend_equivalent(); 1271 if (2 == Mode) OrderAccess::storeload(); 1272 if (!jSelf->is_external_suspend()) return false; 1273 // We raced a suspension -- fall thru into the slow path 1274 TEVENT(ExitSuspendEquivalent - raced); 1275 jSelf->set_suspend_equivalent(); 1276 } 1277 return jSelf->handle_special_suspend_equivalent_condition(); 1278} 1279 1280 1281void ObjectMonitor::ExitEpilog (Thread * Self, ObjectWaiter * Wakee) { 1282 assert(_owner == Self, "invariant"); 1283 1284 // Exit protocol: 1285 // 1. ST _succ = wakee 1286 // 2. membar #loadstore|#storestore; 1287 // 2. ST _owner = NULL 1288 // 3. unpark(wakee) 1289 1290 _succ = Knob_SuccEnabled ? Wakee->_thread : NULL; 1291 ParkEvent * Trigger = Wakee->_event; 1292 1293 // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again. 1294 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be 1295 // out-of-scope (non-extant). 1296 Wakee = NULL; 1297 1298 // Drop the lock 1299 OrderAccess::release_store_ptr(&_owner, NULL); 1300 OrderAccess::fence(); // ST _owner vs LD in unpark() 1301 1302 if (SafepointSynchronize::do_call_back()) { 1303 TEVENT(unpark before SAFEPOINT); 1304 } 1305 1306 DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self); 1307 Trigger->unpark(); 1308 1309 // Maintain stats and report events to JVMTI 1310 if (ObjectMonitor::_sync_Parks != NULL) { 1311 ObjectMonitor::_sync_Parks->inc(); 1312 } 1313} 1314 1315 1316// ----------------------------------------------------------------------------- 1317// Class Loader deadlock handling. 1318// 1319// complete_exit exits a lock returning recursion count 1320// complete_exit/reenter operate as a wait without waiting 1321// complete_exit requires an inflated monitor 1322// The _owner field is not always the Thread addr even with an 1323// inflated monitor, e.g. the monitor can be inflated by a non-owning 1324// thread due to contention. 1325intptr_t ObjectMonitor::complete_exit(TRAPS) { 1326 Thread * const Self = THREAD; 1327 assert(Self->is_Java_thread(), "Must be Java thread!"); 1328 JavaThread *jt = (JavaThread *)THREAD; 1329 1330 DeferredInitialize(); 1331 1332 if (THREAD != _owner) { 1333 if (THREAD->is_lock_owned ((address)_owner)) { 1334 assert(_recursions == 0, "internal state error"); 1335 _owner = THREAD; /* Convert from basiclock addr to Thread addr */ 1336 _recursions = 0; 1337 OwnerIsThread = 1; 1338 } 1339 } 1340 1341 guarantee(Self == _owner, "complete_exit not owner"); 1342 intptr_t save = _recursions; // record the old recursion count 1343 _recursions = 0; // set the recursion level to be 0 1344 exit(true, Self); // exit the monitor 1345 guarantee(_owner != Self, "invariant"); 1346 return save; 1347} 1348 1349// reenter() enters a lock and sets recursion count 1350// complete_exit/reenter operate as a wait without waiting 1351void ObjectMonitor::reenter(intptr_t recursions, TRAPS) { 1352 Thread * const Self = THREAD; 1353 assert(Self->is_Java_thread(), "Must be Java thread!"); 1354 JavaThread *jt = (JavaThread *)THREAD; 1355 1356 guarantee(_owner != Self, "reenter already owner"); 1357 enter(THREAD); // enter the monitor 1358 guarantee(_recursions == 0, "reenter recursion"); 1359 _recursions = recursions; 1360 return; 1361} 1362 1363 1364// ----------------------------------------------------------------------------- 1365// A macro is used below because there may already be a pending 1366// exception which should not abort the execution of the routines 1367// which use this (which is why we don't put this into check_slow and 1368// call it with a CHECK argument). 1369 1370#define CHECK_OWNER() \ 1371 do { \ 1372 if (THREAD != _owner) { \ 1373 if (THREAD->is_lock_owned((address) _owner)) { \ 1374 _owner = THREAD; /* Convert from basiclock addr to Thread addr */ \ 1375 _recursions = 0; \ 1376 OwnerIsThread = 1; \ 1377 } else { \ 1378 TEVENT(Throw IMSX); \ 1379 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); \ 1380 } \ 1381 } \ 1382 } while (false) 1383 1384// check_slow() is a misnomer. It's called to simply to throw an IMSX exception. 1385// TODO-FIXME: remove check_slow() -- it's likely dead. 1386 1387void ObjectMonitor::check_slow(TRAPS) { 1388 TEVENT(check_slow - throw IMSX); 1389 assert(THREAD != _owner && !THREAD->is_lock_owned((address) _owner), "must not be owner"); 1390 THROW_MSG(vmSymbols::java_lang_IllegalMonitorStateException(), "current thread not owner"); 1391} 1392 1393static int Adjust (volatile int * adr, int dx) { 1394 int v; 1395 for (v = *adr; Atomic::cmpxchg(v + dx, adr, v) != v; v = *adr); 1396 return v; 1397} 1398 1399// helper method for posting a monitor wait event 1400void ObjectMonitor::post_monitor_wait_event(EventJavaMonitorWait* event, 1401 jlong notifier_tid, 1402 jlong timeout, 1403 bool timedout) { 1404 event->set_klass(((oop)this->object())->klass()); 1405 event->set_timeout((TYPE_ULONG)timeout); 1406 event->set_address((TYPE_ADDRESS)(uintptr_t)(this->object_addr())); 1407 event->set_notifier((TYPE_OSTHREAD)notifier_tid); 1408 event->set_timedOut((TYPE_BOOLEAN)timedout); 1409 event->commit(); 1410} 1411 1412// ----------------------------------------------------------------------------- 1413// Wait/Notify/NotifyAll 1414// 1415// Note: a subset of changes to ObjectMonitor::wait() 1416// will need to be replicated in complete_exit above 1417void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) { 1418 Thread * const Self = THREAD; 1419 assert(Self->is_Java_thread(), "Must be Java thread!"); 1420 JavaThread *jt = (JavaThread *)THREAD; 1421 1422 DeferredInitialize(); 1423 1424 // Throw IMSX or IEX. 1425 CHECK_OWNER(); 1426 1427 EventJavaMonitorWait event; 1428 1429 // check for a pending interrupt 1430 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1431 // post monitor waited event. Note that this is past-tense, we are done waiting. 1432 if (JvmtiExport::should_post_monitor_waited()) { 1433 // Note: 'false' parameter is passed here because the 1434 // wait was not timed out due to thread interrupt. 1435 JvmtiExport::post_monitor_waited(jt, this, false); 1436 1437 // In this short circuit of the monitor wait protocol, the 1438 // current thread never drops ownership of the monitor and 1439 // never gets added to the wait queue so the current thread 1440 // cannot be made the successor. This means that the 1441 // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally 1442 // consume an unpark() meant for the ParkEvent associated with 1443 // this ObjectMonitor. 1444 } 1445 if (event.should_commit()) { 1446 post_monitor_wait_event(&event, 0, millis, false); 1447 } 1448 TEVENT(Wait - Throw IEX); 1449 THROW(vmSymbols::java_lang_InterruptedException()); 1450 return; 1451 } 1452 1453 TEVENT(Wait); 1454 1455 assert(Self->_Stalled == 0, "invariant"); 1456 Self->_Stalled = intptr_t(this); 1457 jt->set_current_waiting_monitor(this); 1458 1459 // create a node to be put into the queue 1460 // Critically, after we reset() the event but prior to park(), we must check 1461 // for a pending interrupt. 1462 ObjectWaiter node(Self); 1463 node.TState = ObjectWaiter::TS_WAIT; 1464 Self->_ParkEvent->reset(); 1465 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag 1466 1467 // Enter the waiting queue, which is a circular doubly linked list in this case 1468 // but it could be a priority queue or any data structure. 1469 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only 1470 // by the the owner of the monitor *except* in the case where park() 1471 // returns because of a timeout of interrupt. Contention is exceptionally rare 1472 // so we use a simple spin-lock instead of a heavier-weight blocking lock. 1473 1474 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - add"); 1475 AddWaiter(&node); 1476 Thread::SpinRelease(&_WaitSetLock); 1477 1478 if ((SyncFlags & 4) == 0) { 1479 _Responsible = NULL; 1480 } 1481 intptr_t save = _recursions; // record the old recursion count 1482 _waiters++; // increment the number of waiters 1483 _recursions = 0; // set the recursion level to be 1 1484 exit(true, Self); // exit the monitor 1485 guarantee(_owner != Self, "invariant"); 1486 1487 // The thread is on the WaitSet list - now park() it. 1488 // On MP systems it's conceivable that a brief spin before we park 1489 // could be profitable. 1490 // 1491 // TODO-FIXME: change the following logic to a loop of the form 1492 // while (!timeout && !interrupted && _notified == 0) park() 1493 1494 int ret = OS_OK; 1495 int WasNotified = 0; 1496 { // State transition wrappers 1497 OSThread* osthread = Self->osthread(); 1498 OSThreadWaitState osts(osthread, true); 1499 { 1500 ThreadBlockInVM tbivm(jt); 1501 // Thread is in thread_blocked state and oop access is unsafe. 1502 jt->set_suspend_equivalent(); 1503 1504 if (interruptible && (Thread::is_interrupted(THREAD, false) || HAS_PENDING_EXCEPTION)) { 1505 // Intentionally empty 1506 } else 1507 if (node._notified == 0) { 1508 if (millis <= 0) { 1509 Self->_ParkEvent->park(); 1510 } else { 1511 ret = Self->_ParkEvent->park(millis); 1512 } 1513 } 1514 1515 // were we externally suspended while we were waiting? 1516 if (ExitSuspendEquivalent (jt)) { 1517 // TODO-FIXME: add -- if succ == Self then succ = null. 1518 jt->java_suspend_self(); 1519 } 1520 1521 } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm 1522 1523 1524 // Node may be on the WaitSet, the EntryList (or cxq), or in transition 1525 // from the WaitSet to the EntryList. 1526 // See if we need to remove Node from the WaitSet. 1527 // We use double-checked locking to avoid grabbing _WaitSetLock 1528 // if the thread is not on the wait queue. 1529 // 1530 // Note that we don't need a fence before the fetch of TState. 1531 // In the worst case we'll fetch a old-stale value of TS_WAIT previously 1532 // written by the is thread. (perhaps the fetch might even be satisfied 1533 // by a look-aside into the processor's own store buffer, although given 1534 // the length of the code path between the prior ST and this load that's 1535 // highly unlikely). If the following LD fetches a stale TS_WAIT value 1536 // then we'll acquire the lock and then re-fetch a fresh TState value. 1537 // That is, we fail toward safety. 1538 1539 if (node.TState == ObjectWaiter::TS_WAIT) { 1540 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - unlink"); 1541 if (node.TState == ObjectWaiter::TS_WAIT) { 1542 DequeueSpecificWaiter(&node); // unlink from WaitSet 1543 assert(node._notified == 0, "invariant"); 1544 node.TState = ObjectWaiter::TS_RUN; 1545 } 1546 Thread::SpinRelease(&_WaitSetLock); 1547 } 1548 1549 // The thread is now either on off-list (TS_RUN), 1550 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ). 1551 // The Node's TState variable is stable from the perspective of this thread. 1552 // No other threads will asynchronously modify TState. 1553 guarantee(node.TState != ObjectWaiter::TS_WAIT, "invariant"); 1554 OrderAccess::loadload(); 1555 if (_succ == Self) _succ = NULL; 1556 WasNotified = node._notified; 1557 1558 // Reentry phase -- reacquire the monitor. 1559 // re-enter contended monitor after object.wait(). 1560 // retain OBJECT_WAIT state until re-enter successfully completes 1561 // Thread state is thread_in_vm and oop access is again safe, 1562 // although the raw address of the object may have changed. 1563 // (Don't cache naked oops over safepoints, of course). 1564 1565 // post monitor waited event. Note that this is past-tense, we are done waiting. 1566 if (JvmtiExport::should_post_monitor_waited()) { 1567 JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT); 1568 1569 if (node._notified != 0 && _succ == Self) { 1570 // In this part of the monitor wait-notify-reenter protocol it 1571 // is possible (and normal) for another thread to do a fastpath 1572 // monitor enter-exit while this thread is still trying to get 1573 // to the reenter portion of the protocol. 1574 // 1575 // The ObjectMonitor was notified and the current thread is 1576 // the successor which also means that an unpark() has already 1577 // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can 1578 // consume the unpark() that was done when the successor was 1579 // set because the same ParkEvent is shared between Java 1580 // monitors and JVM/TI RawMonitors (for now). 1581 // 1582 // We redo the unpark() to ensure forward progress, i.e., we 1583 // don't want all pending threads hanging (parked) with none 1584 // entering the unlocked monitor. 1585 node._event->unpark(); 1586 } 1587 } 1588 1589 if (event.should_commit()) { 1590 post_monitor_wait_event(&event, node._notifier_tid, millis, ret == OS_TIMEOUT); 1591 } 1592 1593 OrderAccess::fence(); 1594 1595 assert(Self->_Stalled != 0, "invariant"); 1596 Self->_Stalled = 0; 1597 1598 assert(_owner != Self, "invariant"); 1599 ObjectWaiter::TStates v = node.TState; 1600 if (v == ObjectWaiter::TS_RUN) { 1601 enter(Self); 1602 } else { 1603 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant"); 1604 ReenterI(Self, &node); 1605 node.wait_reenter_end(this); 1606 } 1607 1608 // Self has reacquired the lock. 1609 // Lifecycle - the node representing Self must not appear on any queues. 1610 // Node is about to go out-of-scope, but even if it were immortal we wouldn't 1611 // want residual elements associated with this thread left on any lists. 1612 guarantee(node.TState == ObjectWaiter::TS_RUN, "invariant"); 1613 assert(_owner == Self, "invariant"); 1614 assert(_succ != Self , "invariant"); 1615 } // OSThreadWaitState() 1616 1617 jt->set_current_waiting_monitor(NULL); 1618 1619 guarantee(_recursions == 0, "invariant"); 1620 _recursions = save; // restore the old recursion count 1621 _waiters--; // decrement the number of waiters 1622 1623 // Verify a few postconditions 1624 assert(_owner == Self , "invariant"); 1625 assert(_succ != Self , "invariant"); 1626 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant"); 1627 1628 if (SyncFlags & 32) { 1629 OrderAccess::fence(); 1630 } 1631 1632 // check if the notification happened 1633 if (!WasNotified) { 1634 // no, it could be timeout or Thread.interrupt() or both 1635 // check for interrupt event, otherwise it is timeout 1636 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { 1637 TEVENT(Wait - throw IEX from epilog); 1638 THROW(vmSymbols::java_lang_InterruptedException()); 1639 } 1640 } 1641 1642 // NOTE: Spurious wake up will be consider as timeout. 1643 // Monitor notify has precedence over thread interrupt. 1644} 1645 1646 1647// Consider: 1648// If the lock is cool (cxq == null && succ == null) and we're on an MP system 1649// then instead of transferring a thread from the WaitSet to the EntryList 1650// we might just dequeue a thread from the WaitSet and directly unpark() it. 1651 1652void ObjectMonitor::notify(TRAPS) { 1653 CHECK_OWNER(); 1654 if (_WaitSet == NULL) { 1655 TEVENT(Empty-Notify); 1656 return; 1657 } 1658 DTRACE_MONITOR_PROBE(notify, this, object(), THREAD); 1659 1660 int Policy = Knob_MoveNotifyee; 1661 1662 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - notify"); 1663 ObjectWaiter * iterator = DequeueWaiter(); 1664 if (iterator != NULL) { 1665 TEVENT(Notify1 - Transfer); 1666 guarantee(iterator->TState == ObjectWaiter::TS_WAIT, "invariant"); 1667 guarantee(iterator->_notified == 0, "invariant"); 1668 if (Policy != 4) { 1669 iterator->TState = ObjectWaiter::TS_ENTER; 1670 } 1671 iterator->_notified = 1; 1672 Thread * Self = THREAD; 1673 iterator->_notifier_tid = Self->osthread()->thread_id(); 1674 1675 ObjectWaiter * List = _EntryList; 1676 if (List != NULL) { 1677 assert(List->_prev == NULL, "invariant"); 1678 assert(List->TState == ObjectWaiter::TS_ENTER, "invariant"); 1679 assert(List != iterator, "invariant"); 1680 } 1681 1682 if (Policy == 0) { // prepend to EntryList 1683 if (List == NULL) { 1684 iterator->_next = iterator->_prev = NULL; 1685 _EntryList = iterator; 1686 } else { 1687 List->_prev = iterator; 1688 iterator->_next = List; 1689 iterator->_prev = NULL; 1690 _EntryList = iterator; 1691 } 1692 } else 1693 if (Policy == 1) { // append to EntryList 1694 if (List == NULL) { 1695 iterator->_next = iterator->_prev = NULL; 1696 _EntryList = iterator; 1697 } else { 1698 // CONSIDER: finding the tail currently requires a linear-time walk of 1699 // the EntryList. We can make tail access constant-time by converting to 1700 // a CDLL instead of using our current DLL. 1701 ObjectWaiter * Tail; 1702 for (Tail = List; Tail->_next != NULL; Tail = Tail->_next); 1703 assert(Tail != NULL && Tail->_next == NULL, "invariant"); 1704 Tail->_next = iterator; 1705 iterator->_prev = Tail; 1706 iterator->_next = NULL; 1707 } 1708 } else 1709 if (Policy == 2) { // prepend to cxq 1710 // prepend to cxq 1711 if (List == NULL) { 1712 iterator->_next = iterator->_prev = NULL; 1713 _EntryList = iterator; 1714 } else { 1715 iterator->TState = ObjectWaiter::TS_CXQ; 1716 for (;;) { 1717 ObjectWaiter * Front = _cxq; 1718 iterator->_next = Front; 1719 if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) { 1720 break; 1721 } 1722 } 1723 } 1724 } else 1725 if (Policy == 3) { // append to cxq 1726 iterator->TState = ObjectWaiter::TS_CXQ; 1727 for (;;) { 1728 ObjectWaiter * Tail; 1729 Tail = _cxq; 1730 if (Tail == NULL) { 1731 iterator->_next = NULL; 1732 if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) { 1733 break; 1734 } 1735 } else { 1736 while (Tail->_next != NULL) Tail = Tail->_next; 1737 Tail->_next = iterator; 1738 iterator->_prev = Tail; 1739 iterator->_next = NULL; 1740 break; 1741 } 1742 } 1743 } else { 1744 ParkEvent * ev = iterator->_event; 1745 iterator->TState = ObjectWaiter::TS_RUN; 1746 OrderAccess::fence(); 1747 ev->unpark(); 1748 } 1749 1750 if (Policy < 4) { 1751 iterator->wait_reenter_begin(this); 1752 } 1753 1754 // _WaitSetLock protects the wait queue, not the EntryList. We could 1755 // move the add-to-EntryList operation, above, outside the critical section 1756 // protected by _WaitSetLock. In practice that's not useful. With the 1757 // exception of wait() timeouts and interrupts the monitor owner 1758 // is the only thread that grabs _WaitSetLock. There's almost no contention 1759 // on _WaitSetLock so it's not profitable to reduce the length of the 1760 // critical section. 1761 } 1762 1763 Thread::SpinRelease(&_WaitSetLock); 1764 1765 if (iterator != NULL && ObjectMonitor::_sync_Notifications != NULL) { 1766 ObjectMonitor::_sync_Notifications->inc(); 1767 } 1768} 1769 1770 1771void ObjectMonitor::notifyAll(TRAPS) { 1772 CHECK_OWNER(); 1773 ObjectWaiter* iterator; 1774 if (_WaitSet == NULL) { 1775 TEVENT(Empty-NotifyAll); 1776 return; 1777 } 1778 DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD); 1779 1780 int Policy = Knob_MoveNotifyee; 1781 int Tally = 0; 1782 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - notifyall"); 1783 1784 for (;;) { 1785 iterator = DequeueWaiter(); 1786 if (iterator == NULL) break; 1787 TEVENT(NotifyAll - Transfer1); 1788 ++Tally; 1789 1790 // Disposition - what might we do with iterator ? 1791 // a. add it directly to the EntryList - either tail or head. 1792 // b. push it onto the front of the _cxq. 1793 // For now we use (a). 1794 1795 guarantee(iterator->TState == ObjectWaiter::TS_WAIT, "invariant"); 1796 guarantee(iterator->_notified == 0, "invariant"); 1797 iterator->_notified = 1; 1798 Thread * Self = THREAD; 1799 iterator->_notifier_tid = Self->osthread()->thread_id(); 1800 if (Policy != 4) { 1801 iterator->TState = ObjectWaiter::TS_ENTER; 1802 } 1803 1804 ObjectWaiter * List = _EntryList; 1805 if (List != NULL) { 1806 assert(List->_prev == NULL, "invariant"); 1807 assert(List->TState == ObjectWaiter::TS_ENTER, "invariant"); 1808 assert(List != iterator, "invariant"); 1809 } 1810 1811 if (Policy == 0) { // prepend to EntryList 1812 if (List == NULL) { 1813 iterator->_next = iterator->_prev = NULL; 1814 _EntryList = iterator; 1815 } else { 1816 List->_prev = iterator; 1817 iterator->_next = List; 1818 iterator->_prev = NULL; 1819 _EntryList = iterator; 1820 } 1821 } else 1822 if (Policy == 1) { // append to EntryList 1823 if (List == NULL) { 1824 iterator->_next = iterator->_prev = NULL; 1825 _EntryList = iterator; 1826 } else { 1827 // CONSIDER: finding the tail currently requires a linear-time walk of 1828 // the EntryList. We can make tail access constant-time by converting to 1829 // a CDLL instead of using our current DLL. 1830 ObjectWaiter * Tail; 1831 for (Tail = List; Tail->_next != NULL; Tail = Tail->_next); 1832 assert(Tail != NULL && Tail->_next == NULL, "invariant"); 1833 Tail->_next = iterator; 1834 iterator->_prev = Tail; 1835 iterator->_next = NULL; 1836 } 1837 } else 1838 if (Policy == 2) { // prepend to cxq 1839 // prepend to cxq 1840 iterator->TState = ObjectWaiter::TS_CXQ; 1841 for (;;) { 1842 ObjectWaiter * Front = _cxq; 1843 iterator->_next = Front; 1844 if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) { 1845 break; 1846 } 1847 } 1848 } else 1849 if (Policy == 3) { // append to cxq 1850 iterator->TState = ObjectWaiter::TS_CXQ; 1851 for (;;) { 1852 ObjectWaiter * Tail; 1853 Tail = _cxq; 1854 if (Tail == NULL) { 1855 iterator->_next = NULL; 1856 if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) { 1857 break; 1858 } 1859 } else { 1860 while (Tail->_next != NULL) Tail = Tail->_next; 1861 Tail->_next = iterator; 1862 iterator->_prev = Tail; 1863 iterator->_next = NULL; 1864 break; 1865 } 1866 } 1867 } else { 1868 ParkEvent * ev = iterator->_event; 1869 iterator->TState = ObjectWaiter::TS_RUN; 1870 OrderAccess::fence(); 1871 ev->unpark(); 1872 } 1873 1874 if (Policy < 4) { 1875 iterator->wait_reenter_begin(this); 1876 } 1877 1878 // _WaitSetLock protects the wait queue, not the EntryList. We could 1879 // move the add-to-EntryList operation, above, outside the critical section 1880 // protected by _WaitSetLock. In practice that's not useful. With the 1881 // exception of wait() timeouts and interrupts the monitor owner 1882 // is the only thread that grabs _WaitSetLock. There's almost no contention 1883 // on _WaitSetLock so it's not profitable to reduce the length of the 1884 // critical section. 1885 } 1886 1887 Thread::SpinRelease(&_WaitSetLock); 1888 1889 if (Tally != 0 && ObjectMonitor::_sync_Notifications != NULL) { 1890 ObjectMonitor::_sync_Notifications->inc(Tally); 1891 } 1892} 1893 1894// ----------------------------------------------------------------------------- 1895// Adaptive Spinning Support 1896// 1897// Adaptive spin-then-block - rational spinning 1898// 1899// Note that we spin "globally" on _owner with a classic SMP-polite TATAS 1900// algorithm. On high order SMP systems it would be better to start with 1901// a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH, 1902// a contending thread could enqueue itself on the cxq and then spin locally 1903// on a thread-specific variable such as its ParkEvent._Event flag. 1904// That's left as an exercise for the reader. Note that global spinning is 1905// not problematic on Niagara, as the L2$ serves the interconnect and has both 1906// low latency and massive bandwidth. 1907// 1908// Broadly, we can fix the spin frequency -- that is, the % of contended lock 1909// acquisition attempts where we opt to spin -- at 100% and vary the spin count 1910// (duration) or we can fix the count at approximately the duration of 1911// a context switch and vary the frequency. Of course we could also 1912// vary both satisfying K == Frequency * Duration, where K is adaptive by monitor. 1913// See http://j2se.east/~dice/PERSIST/040824-AdaptiveSpinning.html. 1914// 1915// This implementation varies the duration "D", where D varies with 1916// the success rate of recent spin attempts. (D is capped at approximately 1917// length of a round-trip context switch). The success rate for recent 1918// spin attempts is a good predictor of the success rate of future spin 1919// attempts. The mechanism adapts automatically to varying critical 1920// section length (lock modality), system load and degree of parallelism. 1921// D is maintained per-monitor in _SpinDuration and is initialized 1922// optimistically. Spin frequency is fixed at 100%. 1923// 1924// Note that _SpinDuration is volatile, but we update it without locks 1925// or atomics. The code is designed so that _SpinDuration stays within 1926// a reasonable range even in the presence of races. The arithmetic 1927// operations on _SpinDuration are closed over the domain of legal values, 1928// so at worst a race will install and older but still legal value. 1929// At the very worst this introduces some apparent non-determinism. 1930// We might spin when we shouldn't or vice-versa, but since the spin 1931// count are relatively short, even in the worst case, the effect is harmless. 1932// 1933// Care must be taken that a low "D" value does not become an 1934// an absorbing state. Transient spinning failures -- when spinning 1935// is overall profitable -- should not cause the system to converge 1936// on low "D" values. We want spinning to be stable and predictable 1937// and fairly responsive to change and at the same time we don't want 1938// it to oscillate, become metastable, be "too" non-deterministic, 1939// or converge on or enter undesirable stable absorbing states. 1940// 1941// We implement a feedback-based control system -- using past behavior 1942// to predict future behavior. We face two issues: (a) if the 1943// input signal is random then the spin predictor won't provide optimal 1944// results, and (b) if the signal frequency is too high then the control 1945// system, which has some natural response lag, will "chase" the signal. 1946// (b) can arise from multimodal lock hold times. Transient preemption 1947// can also result in apparent bimodal lock hold times. 1948// Although sub-optimal, neither condition is particularly harmful, as 1949// in the worst-case we'll spin when we shouldn't or vice-versa. 1950// The maximum spin duration is rather short so the failure modes aren't bad. 1951// To be conservative, I've tuned the gain in system to bias toward 1952// _not spinning. Relatedly, the system can sometimes enter a mode where it 1953// "rings" or oscillates between spinning and not spinning. This happens 1954// when spinning is just on the cusp of profitability, however, so the 1955// situation is not dire. The state is benign -- there's no need to add 1956// hysteresis control to damp the transition rate between spinning and 1957// not spinning. 1958// 1959 1960intptr_t ObjectMonitor::SpinCallbackArgument = 0; 1961int (*ObjectMonitor::SpinCallbackFunction)(intptr_t, int) = NULL; 1962 1963// Spinning: Fixed frequency (100%), vary duration 1964 1965 1966int ObjectMonitor::TrySpin_VaryDuration (Thread * Self) { 1967 1968 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning. 1969 int ctr = Knob_FixedSpin; 1970 if (ctr != 0) { 1971 while (--ctr >= 0) { 1972 if (TryLock(Self) > 0) return 1; 1973 SpinPause(); 1974 } 1975 return 0; 1976 } 1977 1978 for (ctr = Knob_PreSpin + 1; --ctr >= 0;) { 1979 if (TryLock(Self) > 0) { 1980 // Increase _SpinDuration ... 1981 // Note that we don't clamp SpinDuration precisely at SpinLimit. 1982 // Raising _SpurDuration to the poverty line is key. 1983 int x = _SpinDuration; 1984 if (x < Knob_SpinLimit) { 1985 if (x < Knob_Poverty) x = Knob_Poverty; 1986 _SpinDuration = x + Knob_BonusB; 1987 } 1988 return 1; 1989 } 1990 SpinPause(); 1991 } 1992 1993 // Admission control - verify preconditions for spinning 1994 // 1995 // We always spin a little bit, just to prevent _SpinDuration == 0 from 1996 // becoming an absorbing state. Put another way, we spin briefly to 1997 // sample, just in case the system load, parallelism, contention, or lock 1998 // modality changed. 1999 // 2000 // Consider the following alternative: 2001 // Periodically set _SpinDuration = _SpinLimit and try a long/full 2002 // spin attempt. "Periodically" might mean after a tally of 2003 // the # of failed spin attempts (or iterations) reaches some threshold. 2004 // This takes us into the realm of 1-out-of-N spinning, where we 2005 // hold the duration constant but vary the frequency. 2006 2007 ctr = _SpinDuration; 2008 if (ctr < Knob_SpinBase) ctr = Knob_SpinBase; 2009 if (ctr <= 0) return 0; 2010 2011 if (Knob_SuccRestrict && _succ != NULL) return 0; 2012 if (Knob_OState && NotRunnable (Self, (Thread *) _owner)) { 2013 TEVENT(Spin abort - notrunnable [TOP]); 2014 return 0; 2015 } 2016 2017 int MaxSpin = Knob_MaxSpinners; 2018 if (MaxSpin >= 0) { 2019 if (_Spinner > MaxSpin) { 2020 TEVENT(Spin abort -- too many spinners); 2021 return 0; 2022 } 2023 // Slightly racy, but benign ... 2024 Adjust(&_Spinner, 1); 2025 } 2026 2027 // We're good to spin ... spin ingress. 2028 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades 2029 // when preparing to LD...CAS _owner, etc and the CAS is likely 2030 // to succeed. 2031 int hits = 0; 2032 int msk = 0; 2033 int caspty = Knob_CASPenalty; 2034 int oxpty = Knob_OXPenalty; 2035 int sss = Knob_SpinSetSucc; 2036 if (sss && _succ == NULL) _succ = Self; 2037 Thread * prv = NULL; 2038 2039 // There are three ways to exit the following loop: 2040 // 1. A successful spin where this thread has acquired the lock. 2041 // 2. Spin failure with prejudice 2042 // 3. Spin failure without prejudice 2043 2044 while (--ctr >= 0) { 2045 2046 // Periodic polling -- Check for pending GC 2047 // Threads may spin while they're unsafe. 2048 // We don't want spinning threads to delay the JVM from reaching 2049 // a stop-the-world safepoint or to steal cycles from GC. 2050 // If we detect a pending safepoint we abort in order that 2051 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b) 2052 // this thread, if safe, doesn't steal cycles from GC. 2053 // This is in keeping with the "no loitering in runtime" rule. 2054 // We periodically check to see if there's a safepoint pending. 2055 if ((ctr & 0xFF) == 0) { 2056 if (SafepointSynchronize::do_call_back()) { 2057 TEVENT(Spin: safepoint); 2058 goto Abort; // abrupt spin egress 2059 } 2060 if (Knob_UsePause & 1) SpinPause(); 2061 2062 int (*scb)(intptr_t,int) = SpinCallbackFunction; 2063 if (hits > 50 && scb != NULL) { 2064 int abend = (*scb)(SpinCallbackArgument, 0); 2065 } 2066 } 2067 2068 if (Knob_UsePause & 2) SpinPause(); 2069 2070 // Exponential back-off ... Stay off the bus to reduce coherency traffic. 2071 // This is useful on classic SMP systems, but is of less utility on 2072 // N1-style CMT platforms. 2073 // 2074 // Trade-off: lock acquisition latency vs coherency bandwidth. 2075 // Lock hold times are typically short. A histogram 2076 // of successful spin attempts shows that we usually acquire 2077 // the lock early in the spin. That suggests we want to 2078 // sample _owner frequently in the early phase of the spin, 2079 // but then back-off and sample less frequently as the spin 2080 // progresses. The back-off makes a good citizen on SMP big 2081 // SMP systems. Oversampling _owner can consume excessive 2082 // coherency bandwidth. Relatedly, if we _oversample _owner we 2083 // can inadvertently interfere with the the ST m->owner=null. 2084 // executed by the lock owner. 2085 if (ctr & msk) continue; 2086 ++hits; 2087 if ((hits & 0xF) == 0) { 2088 // The 0xF, above, corresponds to the exponent. 2089 // Consider: (msk+1)|msk 2090 msk = ((msk << 2)|3) & BackOffMask; 2091 } 2092 2093 // Probe _owner with TATAS 2094 // If this thread observes the monitor transition or flicker 2095 // from locked to unlocked to locked, then the odds that this 2096 // thread will acquire the lock in this spin attempt go down 2097 // considerably. The same argument applies if the CAS fails 2098 // or if we observe _owner change from one non-null value to 2099 // another non-null value. In such cases we might abort 2100 // the spin without prejudice or apply a "penalty" to the 2101 // spin count-down variable "ctr", reducing it by 100, say. 2102 2103 Thread * ox = (Thread *) _owner; 2104 if (ox == NULL) { 2105 ox = (Thread *) Atomic::cmpxchg_ptr(Self, &_owner, NULL); 2106 if (ox == NULL) { 2107 // The CAS succeeded -- this thread acquired ownership 2108 // Take care of some bookkeeping to exit spin state. 2109 if (sss && _succ == Self) { 2110 _succ = NULL; 2111 } 2112 if (MaxSpin > 0) Adjust(&_Spinner, -1); 2113 2114 // Increase _SpinDuration : 2115 // The spin was successful (profitable) so we tend toward 2116 // longer spin attempts in the future. 2117 // CONSIDER: factor "ctr" into the _SpinDuration adjustment. 2118 // If we acquired the lock early in the spin cycle it 2119 // makes sense to increase _SpinDuration proportionally. 2120 // Note that we don't clamp SpinDuration precisely at SpinLimit. 2121 int x = _SpinDuration; 2122 if (x < Knob_SpinLimit) { 2123 if (x < Knob_Poverty) x = Knob_Poverty; 2124 _SpinDuration = x + Knob_Bonus; 2125 } 2126 return 1; 2127 } 2128 2129 // The CAS failed ... we can take any of the following actions: 2130 // * penalize: ctr -= Knob_CASPenalty 2131 // * exit spin with prejudice -- goto Abort; 2132 // * exit spin without prejudice. 2133 // * Since CAS is high-latency, retry again immediately. 2134 prv = ox; 2135 TEVENT(Spin: cas failed); 2136 if (caspty == -2) break; 2137 if (caspty == -1) goto Abort; 2138 ctr -= caspty; 2139 continue; 2140 } 2141 2142 // Did lock ownership change hands ? 2143 if (ox != prv && prv != NULL) { 2144 TEVENT(spin: Owner changed) 2145 if (oxpty == -2) break; 2146 if (oxpty == -1) goto Abort; 2147 ctr -= oxpty; 2148 } 2149 prv = ox; 2150 2151 // Abort the spin if the owner is not executing. 2152 // The owner must be executing in order to drop the lock. 2153 // Spinning while the owner is OFFPROC is idiocy. 2154 // Consider: ctr -= RunnablePenalty ; 2155 if (Knob_OState && NotRunnable (Self, ox)) { 2156 TEVENT(Spin abort - notrunnable); 2157 goto Abort; 2158 } 2159 if (sss && _succ == NULL) _succ = Self; 2160 } 2161 2162 // Spin failed with prejudice -- reduce _SpinDuration. 2163 // TODO: Use an AIMD-like policy to adjust _SpinDuration. 2164 // AIMD is globally stable. 2165 TEVENT(Spin failure); 2166 { 2167 int x = _SpinDuration; 2168 if (x > 0) { 2169 // Consider an AIMD scheme like: x -= (x >> 3) + 100 2170 // This is globally sample and tends to damp the response. 2171 x -= Knob_Penalty; 2172 if (x < 0) x = 0; 2173 _SpinDuration = x; 2174 } 2175 } 2176 2177 Abort: 2178 if (MaxSpin >= 0) Adjust(&_Spinner, -1); 2179 if (sss && _succ == Self) { 2180 _succ = NULL; 2181 // Invariant: after setting succ=null a contending thread 2182 // must recheck-retry _owner before parking. This usually happens 2183 // in the normal usage of TrySpin(), but it's safest 2184 // to make TrySpin() as foolproof as possible. 2185 OrderAccess::fence(); 2186 if (TryLock(Self) > 0) return 1; 2187 } 2188 return 0; 2189} 2190 2191// NotRunnable() -- informed spinning 2192// 2193// Don't bother spinning if the owner is not eligible to drop the lock. 2194// Peek at the owner's schedctl.sc_state and Thread._thread_values and 2195// spin only if the owner thread is _thread_in_Java or _thread_in_vm. 2196// The thread must be runnable in order to drop the lock in timely fashion. 2197// If the _owner is not runnable then spinning will not likely be 2198// successful (profitable). 2199// 2200// Beware -- the thread referenced by _owner could have died 2201// so a simply fetch from _owner->_thread_state might trap. 2202// Instead, we use SafeFetchXX() to safely LD _owner->_thread_state. 2203// Because of the lifecycle issues the schedctl and _thread_state values 2204// observed by NotRunnable() might be garbage. NotRunnable must 2205// tolerate this and consider the observed _thread_state value 2206// as advisory. 2207// 2208// Beware too, that _owner is sometimes a BasicLock address and sometimes 2209// a thread pointer. We differentiate the two cases with OwnerIsThread. 2210// Alternately, we might tag the type (thread pointer vs basiclock pointer) 2211// with the LSB of _owner. Another option would be to probablistically probe 2212// the putative _owner->TypeTag value. 2213// 2214// Checking _thread_state isn't perfect. Even if the thread is 2215// in_java it might be blocked on a page-fault or have been preempted 2216// and sitting on a ready/dispatch queue. _thread state in conjunction 2217// with schedctl.sc_state gives us a good picture of what the 2218// thread is doing, however. 2219// 2220// TODO: check schedctl.sc_state. 2221// We'll need to use SafeFetch32() to read from the schedctl block. 2222// See RFE #5004247 and http://sac.sfbay.sun.com/Archives/CaseLog/arc/PSARC/2005/351/ 2223// 2224// The return value from NotRunnable() is *advisory* -- the 2225// result is based on sampling and is not necessarily coherent. 2226// The caller must tolerate false-negative and false-positive errors. 2227// Spinning, in general, is probabilistic anyway. 2228 2229 2230int ObjectMonitor::NotRunnable (Thread * Self, Thread * ox) { 2231 // Check either OwnerIsThread or ox->TypeTag == 2BAD. 2232 if (!OwnerIsThread) return 0; 2233 2234 if (ox == NULL) return 0; 2235 2236 // Avoid transitive spinning ... 2237 // Say T1 spins or blocks trying to acquire L. T1._Stalled is set to L. 2238 // Immediately after T1 acquires L it's possible that T2, also 2239 // spinning on L, will see L.Owner=T1 and T1._Stalled=L. 2240 // This occurs transiently after T1 acquired L but before 2241 // T1 managed to clear T1.Stalled. T2 does not need to abort 2242 // its spin in this circumstance. 2243 intptr_t BlockedOn = SafeFetchN((intptr_t *) &ox->_Stalled, intptr_t(1)); 2244 2245 if (BlockedOn == 1) return 1; 2246 if (BlockedOn != 0) { 2247 return BlockedOn != intptr_t(this) && _owner == ox; 2248 } 2249 2250 assert(sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant"); 2251 int jst = SafeFetch32((int *) &((JavaThread *) ox)->_thread_state, -1);; 2252 // consider also: jst != _thread_in_Java -- but that's overspecific. 2253 return jst == _thread_blocked || jst == _thread_in_native; 2254} 2255 2256 2257// ----------------------------------------------------------------------------- 2258// WaitSet management ... 2259 2260ObjectWaiter::ObjectWaiter(Thread* thread) { 2261 _next = NULL; 2262 _prev = NULL; 2263 _notified = 0; 2264 TState = TS_RUN; 2265 _thread = thread; 2266 _event = thread->_ParkEvent; 2267 _active = false; 2268 assert(_event != NULL, "invariant"); 2269} 2270 2271void ObjectWaiter::wait_reenter_begin(ObjectMonitor *mon) { 2272 JavaThread *jt = (JavaThread *)this->_thread; 2273 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon); 2274} 2275 2276void ObjectWaiter::wait_reenter_end(ObjectMonitor *mon) { 2277 JavaThread *jt = (JavaThread *)this->_thread; 2278 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active); 2279} 2280 2281inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) { 2282 assert(node != NULL, "should not dequeue NULL node"); 2283 assert(node->_prev == NULL, "node already in list"); 2284 assert(node->_next == NULL, "node already in list"); 2285 // put node at end of queue (circular doubly linked list) 2286 if (_WaitSet == NULL) { 2287 _WaitSet = node; 2288 node->_prev = node; 2289 node->_next = node; 2290 } else { 2291 ObjectWaiter* head = _WaitSet; 2292 ObjectWaiter* tail = head->_prev; 2293 assert(tail->_next == head, "invariant check"); 2294 tail->_next = node; 2295 head->_prev = node; 2296 node->_next = head; 2297 node->_prev = tail; 2298 } 2299} 2300 2301inline ObjectWaiter* ObjectMonitor::DequeueWaiter() { 2302 // dequeue the very first waiter 2303 ObjectWaiter* waiter = _WaitSet; 2304 if (waiter) { 2305 DequeueSpecificWaiter(waiter); 2306 } 2307 return waiter; 2308} 2309 2310inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) { 2311 assert(node != NULL, "should not dequeue NULL node"); 2312 assert(node->_prev != NULL, "node already removed from list"); 2313 assert(node->_next != NULL, "node already removed from list"); 2314 // when the waiter has woken up because of interrupt, 2315 // timeout or other spurious wake-up, dequeue the 2316 // waiter from waiting list 2317 ObjectWaiter* next = node->_next; 2318 if (next == node) { 2319 assert(node->_prev == node, "invariant check"); 2320 _WaitSet = NULL; 2321 } else { 2322 ObjectWaiter* prev = node->_prev; 2323 assert(prev->_next == node, "invariant check"); 2324 assert(next->_prev == node, "invariant check"); 2325 next->_prev = prev; 2326 prev->_next = next; 2327 if (_WaitSet == node) { 2328 _WaitSet = next; 2329 } 2330 } 2331 node->_next = NULL; 2332 node->_prev = NULL; 2333} 2334 2335// ----------------------------------------------------------------------------- 2336// PerfData support 2337PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = NULL; 2338PerfCounter * ObjectMonitor::_sync_FutileWakeups = NULL; 2339PerfCounter * ObjectMonitor::_sync_Parks = NULL; 2340PerfCounter * ObjectMonitor::_sync_EmptyNotifications = NULL; 2341PerfCounter * ObjectMonitor::_sync_Notifications = NULL; 2342PerfCounter * ObjectMonitor::_sync_PrivateA = NULL; 2343PerfCounter * ObjectMonitor::_sync_PrivateB = NULL; 2344PerfCounter * ObjectMonitor::_sync_SlowExit = NULL; 2345PerfCounter * ObjectMonitor::_sync_SlowEnter = NULL; 2346PerfCounter * ObjectMonitor::_sync_SlowNotify = NULL; 2347PerfCounter * ObjectMonitor::_sync_SlowNotifyAll = NULL; 2348PerfCounter * ObjectMonitor::_sync_FailedSpins = NULL; 2349PerfCounter * ObjectMonitor::_sync_SuccessfulSpins = NULL; 2350PerfCounter * ObjectMonitor::_sync_MonInCirculation = NULL; 2351PerfCounter * ObjectMonitor::_sync_MonScavenged = NULL; 2352PerfCounter * ObjectMonitor::_sync_Inflations = NULL; 2353PerfCounter * ObjectMonitor::_sync_Deflations = NULL; 2354PerfLongVariable * ObjectMonitor::_sync_MonExtant = NULL; 2355 2356// One-shot global initialization for the sync subsystem. 2357// We could also defer initialization and initialize on-demand 2358// the first time we call inflate(). Initialization would 2359// be protected - like so many things - by the MonitorCache_lock. 2360 2361void ObjectMonitor::Initialize() { 2362 static int InitializationCompleted = 0; 2363 assert(InitializationCompleted == 0, "invariant"); 2364 InitializationCompleted = 1; 2365 if (UsePerfData) { 2366 EXCEPTION_MARK; 2367 #define NEWPERFCOUNTER(n) {n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events,CHECK); } 2368 #define NEWPERFVARIABLE(n) {n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events,CHECK); } 2369 NEWPERFCOUNTER(_sync_Inflations); 2370 NEWPERFCOUNTER(_sync_Deflations); 2371 NEWPERFCOUNTER(_sync_ContendedLockAttempts); 2372 NEWPERFCOUNTER(_sync_FutileWakeups); 2373 NEWPERFCOUNTER(_sync_Parks); 2374 NEWPERFCOUNTER(_sync_EmptyNotifications); 2375 NEWPERFCOUNTER(_sync_Notifications); 2376 NEWPERFCOUNTER(_sync_SlowEnter); 2377 NEWPERFCOUNTER(_sync_SlowExit); 2378 NEWPERFCOUNTER(_sync_SlowNotify); 2379 NEWPERFCOUNTER(_sync_SlowNotifyAll); 2380 NEWPERFCOUNTER(_sync_FailedSpins); 2381 NEWPERFCOUNTER(_sync_SuccessfulSpins); 2382 NEWPERFCOUNTER(_sync_PrivateA); 2383 NEWPERFCOUNTER(_sync_PrivateB); 2384 NEWPERFCOUNTER(_sync_MonInCirculation); 2385 NEWPERFCOUNTER(_sync_MonScavenged); 2386 NEWPERFVARIABLE(_sync_MonExtant); 2387 #undef NEWPERFCOUNTER 2388 } 2389} 2390 2391 2392// Compile-time asserts 2393// When possible, it's better to catch errors deterministically at 2394// compile-time than at runtime. The down-side to using compile-time 2395// asserts is that error message -- often something about negative array 2396// indices -- is opaque. 2397 2398#define CTASSERT(x) { int tag[1-(2*!(x))]; printf ("Tag @" INTPTR_FORMAT "\n", (intptr_t)tag); } 2399 2400void ObjectMonitor::ctAsserts() { 2401 CTASSERT(offset_of (ObjectMonitor, _header) == 0); 2402} 2403 2404 2405static char * kvGet (char * kvList, const char * Key) { 2406 if (kvList == NULL) return NULL; 2407 size_t n = strlen(Key); 2408 char * Search; 2409 for (Search = kvList; *Search; Search += strlen(Search) + 1) { 2410 if (strncmp (Search, Key, n) == 0) { 2411 if (Search[n] == '=') return Search + n + 1; 2412 if (Search[n] == 0) return(char *) "1"; 2413 } 2414 } 2415 return NULL; 2416} 2417 2418static int kvGetInt (char * kvList, const char * Key, int Default) { 2419 char * v = kvGet(kvList, Key); 2420 int rslt = v ? ::strtol(v, NULL, 0) : Default; 2421 if (Knob_ReportSettings && v != NULL) { 2422 ::printf (" SyncKnob: %s %d(%d)\n", Key, rslt, Default) ; 2423 ::fflush(stdout); 2424 } 2425 return rslt; 2426} 2427 2428void ObjectMonitor::DeferredInitialize() { 2429 if (InitDone > 0) return; 2430 if (Atomic::cmpxchg (-1, &InitDone, 0) != 0) { 2431 while (InitDone != 1); 2432 return; 2433 } 2434 2435 // One-shot global initialization ... 2436 // The initialization is idempotent, so we don't need locks. 2437 // In the future consider doing this via os::init_2(). 2438 // SyncKnobs consist of <Key>=<Value> pairs in the style 2439 // of environment variables. Start by converting ':' to NUL. 2440 2441 if (SyncKnobs == NULL) SyncKnobs = ""; 2442 2443 size_t sz = strlen(SyncKnobs); 2444 char * knobs = (char *) malloc(sz + 2); 2445 if (knobs == NULL) { 2446 vm_exit_out_of_memory(sz + 2, OOM_MALLOC_ERROR, "Parse SyncKnobs"); 2447 guarantee(0, "invariant"); 2448 } 2449 strcpy(knobs, SyncKnobs); 2450 knobs[sz+1] = 0; 2451 for (char * p = knobs; *p; p++) { 2452 if (*p == ':') *p = 0; 2453 } 2454 2455 #define SETKNOB(x) { Knob_##x = kvGetInt (knobs, #x, Knob_##x); } 2456 SETKNOB(ReportSettings); 2457 SETKNOB(Verbose); 2458 SETKNOB(FixedSpin); 2459 SETKNOB(SpinLimit); 2460 SETKNOB(SpinBase); 2461 SETKNOB(SpinBackOff); 2462 SETKNOB(CASPenalty); 2463 SETKNOB(OXPenalty); 2464 SETKNOB(LogSpins); 2465 SETKNOB(SpinSetSucc); 2466 SETKNOB(SuccEnabled); 2467 SETKNOB(SuccRestrict); 2468 SETKNOB(Penalty); 2469 SETKNOB(Bonus); 2470 SETKNOB(BonusB); 2471 SETKNOB(Poverty); 2472 SETKNOB(SpinAfterFutile); 2473 SETKNOB(UsePause); 2474 SETKNOB(SpinEarly); 2475 SETKNOB(OState); 2476 SETKNOB(MaxSpinners); 2477 SETKNOB(PreSpin); 2478 SETKNOB(ExitPolicy); 2479 SETKNOB(QMode); 2480 SETKNOB(ResetEvent); 2481 SETKNOB(MoveNotifyee); 2482 SETKNOB(FastHSSEC); 2483 #undef SETKNOB 2484 2485 if (os::is_MP()) { 2486 BackOffMask = (1 << Knob_SpinBackOff) - 1; 2487 if (Knob_ReportSettings) ::printf("BackOffMask=%X\n", BackOffMask); 2488 // CONSIDER: BackOffMask = ROUNDUP_NEXT_POWER2 (ncpus-1) 2489 } else { 2490 Knob_SpinLimit = 0; 2491 Knob_SpinBase = 0; 2492 Knob_PreSpin = 0; 2493 Knob_FixedSpin = -1; 2494 } 2495 2496 if (Knob_LogSpins == 0) { 2497 ObjectMonitor::_sync_FailedSpins = NULL; 2498 } 2499 2500 free(knobs); 2501 OrderAccess::fence(); 2502 InitDone = 1; 2503} 2504 2505#ifndef PRODUCT 2506void ObjectMonitor::verify() { 2507} 2508 2509void ObjectMonitor::print() { 2510} 2511#endif 2512